PACIFIC JOURNAL OF MATHEMATICS Vol. 81, No. 1, 1979 ON THE AVERAGE NUMBER OF REAL ZEROS OF A CLASS OF RANDOM ALGEBRAIC CURVES M. SAMBANDHAM Let a lf α 2,, be a sequence of dependent normal rom variables with mean zero, variance one the correlation between any two rom variables is p, 0<p<l. In this paper the average number of real zeros of Σ*=i a k k p x k, 0^p<oo is estimated for large n this average is asymptotic to (2π)- 1 [l+(2p+l) 1/2 logn. 1* Let a 19 a 2, be a sequence of dependent normal rom variables with mean zero, variance one joint density function. (1.1) I M\ 1/2 (2π)- n/2 exp [-(l/2)α'λγa] where Af~ ] is the moment matrix with p tj p, i Φ j, 0 < p < 1, i, j 1, 2, '',n. We estimate in this paper the average number of real zeros of (1.2) f{x) = Σ ^x k, 0 ^ p < - we state our result in the following theorem. THEOREM. The average number of real zeros of (1.2) in co<; x <^ co, when the rom variables are dependent normal with joint density function (1.1) is (2π)" 1 [l + (2p + l) 1/2 ] log n 9 for larger n. When p = 0, that is, for the polynomial Σ a tp*9 the average number of real zeros is estimated in Sambham [5] this average is π" 1 log n. Since the maxima or minima of Σ a^k is only half of the average number of real zeros of Σ ha k x k ~~\ by giving p = 1 in the theorem we get the average number of maxima of Σ a k^k- This average has been already estimated in Sambham Bhatt [6] its value is (4π)- ι [l + 3 1/2 ] log n. When the rom variables are independent normally distributed Das [2] estimated the average number of real zeros of [1.2] this average is π~ ι [l + (2p + 1) 1/2 ] log n. Under the same condition the average number of maxima of Σ a k% k is (2π)~ 1 [l + 3 1/2 ] log n the average number of real zeros of Σ a k χk i s (2/π) log n. These two results are respectively in Das [1] Kac [3]. We note that the average number of zeros the average number of maxima in the case when the rom variables are in dependent are twice that of the case when the rom variables are 207
208 M. SAMBANDHAM dependent normal with a constant correlation. This is because when the rom variables are dependent with a constant correlation p, most of the rom variables have a tendency to be of the same sign as they are interdependent. As the most of the rom variables preserve the same sign Σ a k k p x k has a tendency of behaving like ±^\a k \k p x k. Under this condition when x > 0, the consecutive terms have a tendency to cancel each other when x < 0 the cancellation does not become possible. This fact reduces the average number of real zeros for x > 0 to o(log n). In view of the relation f{x) = tt*s +1 Σ <* -*(! ~ kn-yy k+1 & = 0 == n p x n+1 P n (y), y = x the number of roots of in ( oo, 1) U (1, ) equals with probability one, the number of roots of the polynomial P n (y) in ( 1, 1). Proceeding the method here we can easily show that the number of zeros of the polynomial ^t^oa k x k in ( 1,1) remain true for P n (y) in ( 1, 1). Hence we get from Sambham [5] (1.3) Λf Λ (l, oo) = o(logn) (1.4) M n (- oo, 1) ~ (27Γ)" 1 logn. Therefore our further discussion will be on the average number of real zeros of (1.2) in ( 1, 1). If we show that (1.5) M % {- 1, 0) ~ (2π)-\2p + 1) 1/2 log n (1.6) M n (0, 1) = o(log n) in view of the relations (1.3) (1.4) we get the proof of the theorem. To prove (1.5) (1.6) we proceed as follows: 2* Let M n (a, b) denote the average number of real zeros of (1.2) in (a, b). Then following the method in Sambham [5] we get (2.1) M n (a, b) - [[(A P C P - where
ON THE AVERAGE NUMBER OF REAL ZEROS OF A CLASS 2Θ9 A p = A p (x) = (1 - p) Σ B p = C p Ξ if ApCp JSp > 0 in (α, 6) which is easily seen to hold as in Sambham [5]. Since I."-te [* (!;')]} I da; L ezαλ 1 a? /J we can sum the values of A p, B p C 9. This calculations show that for large n 0 < a? ^ 1 (log log w/n) - Bj A 9 C P -B* A /"^ ~D 2 A /""* ~D 2 ^ (~1 ~D 2 since each ^ 0 -^-1 Άf> 1 A^C^ - BU (1 - xy Here in the following Zr(a?, p) with subscripts are bounded positive values of x all of them are greater than zero. Therefore we find {A,C,-BW* ^ L ( χ s (A 0 C 0 - B y L β (x, p) Therefore (2.1) reduces to (2.8) AΓ./0,1 - l 0 g l 0 g W ) = 0(1). \ n /
210 M. SAMBANDHAM Since always C 2 V /2 ) (2.4) MA - l o g l o g^ l) = 0(lo* log n) (2.3) (2.4) proves (1 6). Now we proceed to prove (1.5). When 1 <; x <; 0 we find that in A Pf B p C p the first terms in the right h side are dominant in this case we get (A P G P - Btr 2 < (ΛC 0 -.B 02 ) 1/2 L ^ v ) < L 8 (a?, p) ^ yj y4 I /y * Therefore for 1 + ^^α ^O, where ίy = exp [ (log w) 1/3 ] we get (2.5) Λf n (-1 + 57, 0) = 0(log n) 1/s For 1 ^ x ^ 1 + δ/n, where δ (log n) 1/2, we have therefore (2.6) MJ-1, -1 + A) = 0(log w) 1/2. For a? in the interval ( 1 + δ/n, 1 17) we follow the method suggested by Logan Shepp [4], which was used by Das [2] also. 3. We put From Kac [3] we get μ ε (x) = 1 if ε < x < ε = 0 otherwise. (3.1) M n (a, b) = lim(2εr\ b E{μ ε (f(x))\f\x)\}dx. ε-» 0 J α The combined variable (f(x), f\x)) has characteristic function, φ(z, w) 7{exp [i f(x)z + ίf'(x)w]}. The probability density p(ξ, η) for f(x) = ξ f\x) =77 is given by p(f> V) (2 π )\ \ e χ P [ iζz i^w]9>(2, w)dzdw.
ON THE AVERAGE NUMBER OF REAL ZEROS OF A CLASS 211 Therefore the chance that u ^ fix) < u + du u ^ f'(x) < v + dv hold together in p(u, v)du du. As the x's very both / /' assume values from - co to w independently to one another so that Let us write E[μλf)\f'\] = F(u) \ \v\p(u, v)dv. J oo Then F(u) is continuous therefore we get lim(2ε)- 1 E[μ ε (f)\f'\] = lim(2s)- 1 ( 0 J F(u)du we get the Kac Rice formula. S b!*z{f)\f'\dx is bounded from (3.1) α M n (a, b) = lim(2ε)\ b E[μ ε (f)\f'\]dx (3.2) = \ b F(0)dx = Γ dx ( \η\p(fi, η)dη. Ja Ja J -co We put f(x) = Σ'Li α Λ 6fc f\x) = Σϊ=i «^7ί so that 9>(s, w) - exp [( - y)^ 1-1 ) Σ(M + CJCW) 2 (3.3) p(0, i/) = (27r)~ 2 \ dw \ exp (-ίyw)φ(z, w)dz. J -co J oo Then for ε > 0 we have where Re sts for the real part. We need the following identity, valid for non zero P Q,
212 M. SAMBANDHAM (3.5) Re τr 2 Γ dw Γ exp Γ - {±r\pz + Qwf\dz = 0. One way to see this is to allow b k c k to be arbitrary in (3.4). If we take them each to be constant in k then the probability density p(ξ> V) corresponding to ξ = Σ &J>* = P# ^ = Σ α *A = 2/ degenerates (3.5) follows. Further given P Q, the constants x?/ can be chosen such that x y are normally distributed. We choose P Q such that (3.6) p 2 - (1 - p) Σ δϊ PQ = (1 - P) Σ From (3.4) (3.5) we get (3.7) - exp{ - (-ί We put 2 = w'w, w = xw' use Frullani's theorem to integrate on w\ The right h side of (3.7) reduces to (3.8) g n (x) - 2^L l o g hn^xf where K(x, u) _ {[(1 -p) + ρ\]u 2-2[(1 - i)λ 2 + px,]u + [(1 - where λ 2 s λ 3 = χ 3 (χ) = λ 5 Ξ
ON THE AVERAGE NUMBER OF REAL ZEROS OF A CLASS 213 We put a 1 + δ/ n, 6 = 1 + 7, % exp ( t/2ri), u = nv/t. Therefore Af^-1+, -1 + 7J = 1 g n (x] (3.9) = (47rr i Γ 0 vi O logtγ Λ (t, where V»(ί, v) U n (t, v) = [(1 -p) + px n ]v z + [(1 - p)x tl [(1 - /o) + ^jl - (1 ~ ^V + f λ31 T L (1 -!θ) + θλ u J i V ^/J [φ * "»(-?)] ' n
214 M. SAMBANDBAM 4-2 Λ'δlW ' o [Σ(-i [Σ p(~)] From Das [2] we get ^21 = = (2p + 1) λ 41 = = (2p + 1) ^using the idea in (2.2) we get x sι = + 00 <2p + 2) + 0( <Γ ί/2 ) o(±) \ r> 1/2\ lϋ ), 2p + l 2p+3 Now where log W.{v, t)dv = Llog (^2 ~ ^L + ^)(L 2 + 2λL + μ) g * k ; ' K (L 2-2λL + λ 2 )(L 2 + 2XL + λ 2 ) ~ 2 λ L (L 2-2λL + λ 2 )(L 2 + 2XL + λ 2 ) ^ ~ μ)dv λ = λ(t) = (1 - p) + For (l/2)(log w) 1/2 <: t ^ nδ when ^ is large, we find that λ 31, λ n λ 51 are tending to zero, λ 21 λ 41 are respectively asymptotic to (2p + 1) (2p + l)(2p + 2) (3.10) ^r **\ L log ^-2^ + ^ d, = 0 (L logn ). - 2Xv + λ 2 Further we note that
ON THE AVERAGE NUMBER OF REAL ZEROS OF A CLASS 215 for large L. V ύ ΔT V + 8 This makes τ 0 t i-l v 2 2λ?; + μ 2-2), I) 2 Where ^ <εlog?ι ε is infinitely small. obtain from (3.9), (3.10) (3.11) Taking L large we l +,-l + n) = (2π)-\2<p + l) 1/2 logn + o(logn). n I Hence we have proved (1.5) combining this with the discussion in 1 2 we get the proof of the theorem. REFERENCES 1. M. Das, The average number of maxima of a rom algebraic curve, Proc. Camb. Phil. Soc, 65 (1969), 741-53. 2., Real zeros of a class of rom algebraic polynomials, J. Indian Math. Soc, 36 (1972), 53-63. 3. M. Kac, On the average number of real roots of a rom algebraic equation, Bull. Amer. Math. Soc. 49 (1943) 314-20. 4. B. F. Logan L. A. Shepp, Real zeros of rom polynomials II, Proc. Lond. Math. Soc, 18 (1968), 308-14. 5. M. Sambham, On the real roots of the rom algebraic equation. To appear in Indian J. Pure. Appl. Math. 6. M. Sambhan S. S. Bhatt. On the average number of maxima of a rom algebraic curve, Submitted for publication. Received April 20, 1977 A. C. COLLEGE OF ENGINEERING AND TECHNOLOGY KARAIKUDI-623 004, INDIA